Make custom_ops import optional in native runners and improve error message

examples/models/llama/runner/native.py

@@ -23,7 +23,10 @@
 from executorch.examples.models.llama.runner.generation import LlamaRunner
 
 # Note: import this after portable_lib
-from executorch.extension.llm.custom_ops import custom_ops  # noqa # usort: skip
+try:
+    from executorch.extension.llm.custom_ops import custom_ops  # noqa # usort: skip
+except Exception:
+    pass  # Not needed for pybindings inference.
 from executorch.kernels import quantized  # noqa
 
 

examples/models/llama3_2_vision/runner/native.py

@@ -26,7 +26,10 @@
 from executorch.extension.pybindings import portable_lib  # noqa # usort: skip
 
 # Note: import this after portable_lib
-from executorch.extension.llm.custom_ops import custom_ops  # noqa # usort: skip
+try:
+    from executorch.extension.llm.custom_ops import custom_ops  # noqa # usort: skip
+except Exception:
+    pass  # Not needed for pybindings inference.
 from executorch.kernels import quantized  # noqa
 
 

extension/llm/custom_ops/custom_ops.py

@@ -41,7 +41,12 @@
 
     libs = list(package_path.glob("**/*custom_ops_aot_lib.*"))
 
-    assert len(libs) == 1, f"Expected 1 library but got {len(libs)}"
+    assert len(libs) == 1, (
+        f"Expected 1 custom_ops_aot_lib library but got {len(libs)} "
+        f"(searched in {package_path}). "
+        "If building from source, re-build with "
+        "-DEXECUTORCH_BUILD_KERNELS_LLM_AOT=ON or use the pybind cmake preset."
+    )
     logging.info(f"Loading custom ops library: {libs[0]}")
     torch.ops.load_library(libs[0])
     op = torch.ops.llama.sdpa_with_kv_cache.default